The future hinted at in Spike Jonze’s 2013 film “Her” is edging closer to reality, with startups building AI agents that collaborate to fill out forms, call customer service, and order food delivery. Investors from NEA, Mayfield Fund, and Foundation Capital say these “multi-agent systems” are the next big thing.
A multi-agent system is a distributed system of multiple specialized AI agents working together towards a common goal. These agents break down a task into multiple smaller steps, each working on a specific task to achieve a broader goal.
“You give it a goal, and it breaks that goal down into a series of steps and creates a sequence, and then it can execute that sequence,” said Ash Garg of Foundation Capital.
Ash Garg of Foundation Capital Ash Garg
Using the example of planning a trip to Italy, Garg said, “Language models at scale are [presenting] “You can make an itinerary, but that’s not a true vacation plan,” an LLM like OpenAI’s GPT-4 would respond in a question-and-answer format, and likely lack the reasoning power to go through all the steps required to plan a trip.
In contrast, multi-agent systems are built on LLM and are designed to execute a task from start to finish. In the Italy example, the agent system undertakes to plan the entire trip, including booking flights and hotels, reserving restaurants, and recommending activities. Throughout the process, the agents aim to continuously verify the accuracy and completeness of their work.
Madison Faulkner of the NEA said the verification step is crucial for multi-agent systems: By defining success criteria, these systems can autonomously verify whether they are achieving their goals throughout the process.
“This is where we get a really powerful, scalable approach to building agent systems.”
Additionally, AI agents collaborate throughout the process by calling APIs and accessing relevant data. Where necessary, these systems may also include human input, commonly referred to as human-in-the-loop systems.
In the emerging space, there are already many startups operating. Many of them focus on specific use cases at the application layer that general-purpose LLMs cannot adequately address. For example, Regie AI, which has raised $20.8 million according to PitchBook, uses an “autopilot sales agent” to automatically mine leads, create customized emails, and follow up with buyers. Cognition raised $175 million at a $2 billion valuation six months after founding, according to PitchBook. The company is developing Devin, an autonomous AI engineer that can perform complex engineering tasks.
Agent startups “can come up with solutions to very niche or more specific problems that the big AI companies can’t tackle,” said Warren Hui of Seoul Ventures.
Other companies are focusing on the infrastructure layer to power multi-agent systems. Emergence, which has raised $155.9 million according to PitchBook, orchestrates between first- and third-party models to automate enterprise knowledge work like report summarization and claims processing.
AI Startup Emergence Satya Nitta
“We’re trying to bring some deep thinking to the table about what that means. [for agents] “The agents are self-improving,” says Emergence CEO Satya Nitta, and the goal is to “give agents the ability to control and operate any kind of software or computer system in a human-like way.”
Another startup, Phidata, is turning LLMs into AI assistants by enabling real-time data access, including web searches and database queries. “Our orders are skyrocketing,” says Phidata CEO Ashpreet Bedi. “We’ve never seen a price push back.”
However, there are also companies that span both the LLM and agent layers. One example is Sakana AI, a Japan-based startup that has raised $155 million, according to PitchBook. Sakana AI is a large foundational model with elements of a multi-agent system, according to Faulkner. The large model breaks down expertise into smaller specialized models, focusing on different domains such as Japanese language and mathematics.
Another example is Hippocratic AI, a healthcare-focused foundation that has raised $120 million, according to PitchBook, and is partnering with Nvidia to build a voice-based AI agent for patient-facing tasks, according to its website.
As more startups enter the space and investors pour in money, Her’s vision seems closer than ever. “For the first time in history, you can have agents at your disposal to automate multiple workflows,” Nitta said.
Navin Chaddha of the Mayfield Fund is optimistic: AI agents will work as teammates with employees. “A person like me will be working with 50 to 100 agents — chat agents, support agents, service agents, research agents, the list goes on and on,” Chaddha says. “And those agents will need to talk to each other.”
